feat(phase-3): AC-9~12 后端任务引擎稳定性(队列隔离/心跳降级/清理瘦投影/终态与超时)

- AC-9(S1-B2/B3):视频/图像任务改走独立 videoQueueExecutor(aiimage.task-queue.video-max-concurrent,
  默认 6),不再与结果文件组装共用 12 格 taskQueueExecutor;TaskQueueGate 增 queue 标签区分指标;
  scheduled worker 本地派发被拒改回队(requeue),不再压回 16 线程调度池内联组装。
- AC-10(S1-B4):CollectDataService.updateProgress 去掉任务锁,改无条件短事务 + updated_at 乐观 CAS,
  只写 result_json/updated_at 两列;TaskHeartbeatService 对 SPI 进度写异常降级,
  TaskHeartbeatVo 新增 progressDegraded/warning,alive 仍为 true。
- AC-11(S1-B5/B6):ModuleHistoryCleanupService 批查与四路指针查询全部瘦投影;
  application.yml 增 server.shutdown=graceful 与 spring.lifecycle.timeout-per-shutdown-phase=30s;
  虚拟线程池 destroy 改为 await 在途任务(上限 aiimage.task-queue.shutdown-await-seconds)。
- AC-12(S1-B1/B7/B8/B9):PENDING 僵尸重试计数可到 MAX 并就地翻 FAILED 进 exhaustedJobs;
  修正 selectAdminTaskDetail 的双转义(>= -> >=)并按真实 SQL 修正迁移基准账本;
  MyBatis default-statement-timeout=60s(AIIMAGE_DB_DEFAULT_STATEMENT_TIMEOUT_SECONDS);
  店铺数据两处删除收进 @Transactional,对象回收沿用 afterCommit 注册。
- 契约激活:PendingPhase3*Test 14 例全部去 @Disabled;同步改写被新契约反转的既有判据
  (CollectDataProgressThrottleTest 锁/updateById 口径、OwnerRouting 的内联分支改显式关闭派发形态、
  FaultInjection 指标按新 queue 标签查找)与三处棘轮(catch total 1031->1034、
  notRethrow 582->585、超管调用点 17->18,均逐条记名)。
This commit is contained in:
2026-10-28 21:05:33 +08:00
parent c4fc9784e8
commit 62a1b50067
24 changed files with 338 additions and 113 deletions
@@ -109,6 +109,13 @@ AIIMAGE_RESULT_FILE_JOB_LOCAL_WORKER_ENABLED=false
AIIMAGE_RESULT_FILE_JOB_LOCAL_WORKER_DELAY_MS=10000
AIIMAGE_RESULT_FILE_JOB_STUCK_TIMEOUT_MINUTES=5
# 任务队列(可选,留空即用 application.yml 的出厂默认值):
# 视频/图像 Coze 任务与结果文件组装各自独立许可;优雅停机等待在途任务的秒数。
AIIMAGE_TASK_QUEUE_VIDEO_MAX_CONCURRENT=6
AIIMAGE_TASK_QUEUE_VIDEO_MAX_WAITING=1000
AIIMAGE_TASK_QUEUE_SHUTDOWN_AWAIT_SECONDS=30
AIIMAGE_DB_DEFAULT_STATEMENT_TIMEOUT_SECONDS=60
AIIMAGE_ZINIAO_ENABLED=false
AIIMAGE_ZINIAO_BASE_URL=https://sbappstoreapi.ziniao.com/openapi-router
AIIMAGE_ZINIAO_API_KEY=change-me
@@ -1,9 +1,14 @@
server:
port: ${AIIMAGE_SERVER_PORT:18080}
# S1-B6:发版硬切在途上传/组装的根因。graceful 让 Tomcat 先停止接收新请求、
# 等在途请求收尾;等待上限由 spring.lifecycle.timeout-per-shutdown-phase 控制。
shutdown: graceful
spring:
application:
name: aiimage-backend
lifecycle:
timeout-per-shutdown-phase: 30s
servlet:
multipart:
max-file-size: 2GB
@@ -82,6 +87,9 @@ mybatis-plus:
mapper-locations: classpath*:mapper/**/*.xml
configuration:
map-underscore-to-camel-case: true
# S1-B8:语句级超时(秒)。socketTimeout=300s + 30 格池 ⇒ 一条慢 SQL 能占死一格连接 5 分钟,
# Tomcat 请求在 30s 取连接处失败。60s 覆盖现网最慢的报表/清理语句,又留足连接回收余量。
default-statement-timeout: ${AIIMAGE_DB_DEFAULT_STATEMENT_TIMEOUT_SECONDS:60}
log-impl: org.apache.ibatis.logging.slf4j.Slf4jImpl
springdoc:
@@ -382,6 +390,11 @@ aiimage:
task-queue:
max-concurrent: ${AIIMAGE_TASK_QUEUE_MAX_CONCURRENT:12}
max-waiting: ${AIIMAGE_TASK_QUEUE_MAX_WAITING:1000}
# S1-B2:图像/视频(Coze 读超时 1 小时)独立许可数,不与结果文件组装的 12 格共用。
video-max-concurrent: ${AIIMAGE_TASK_QUEUE_VIDEO_MAX_CONCURRENT:6}
video-max-waiting: ${AIIMAGE_TASK_QUEUE_VIDEO_MAX_WAITING:1000}
# S1-B6:优雅停机时虚拟线程池等待在途任务的秒数(对齐 spring.lifecycle 的 30s)。
shutdown-await-seconds: ${AIIMAGE_TASK_QUEUE_SHUTDOWN_AWAIT_SECONDS:30}
brand-check:
base-url: ${AIIMAGE_BRAND_CHECK_BASE_URL:http://47.110.241.161:16890}
path: ${AIIMAGE_BRAND_CHECK_PATH:/brand_check}
@@ -37,7 +37,7 @@
FROM biz_image_video_async_task t
LEFT JOIN users u ON u.id = t.user_id
WHERE t.task_type = 'IMAGE_VIDEO_WORKFLOW'
AND t.submitted_at >= DATE_SUB(NOW(), INTERVAL 3 DAY)
AND t.submitted_at >= DATE_SUB(NOW(), INTERVAL 3 DAY)
AND t.id = #{taskId}
LIMIT 1
</select>